EvoSyn: Generalizable Evolutionary Data Synthesis for Verifiable Learning

TL;DR

EvoSyn introduces a general, task-agnostic framework for synthesizing verifiable data by evolving data-filtering strategies with a consistency-based evaluator, transforming data filtering into a principled optimization process. From minimal seed supervision, EvoSyn jointly synthesizes problems, candidate solutions, and tests, producing a coherent verifiable dataset usable for RLVR and model distillation. Empirical results on LiveCodeBench and AgentBench-OS show substantial improvements across multiple models and training paradigms, with ablations highlighting the value of strategy evolution and Zero-Variance Pruning. The work demonstrates robust cross-domain generalization and provides a scalable approach to reducing reliance on handcrafted, domain-specific verification heuristics, albeit with costs related to verification effort and data-diversity challenges.

Abstract

Reliable verifiable data has become a key driver of capability gains in modern language models, enabling stable reinforcement learning with verifiable rewards and effective distillation that transfers competence across math, coding, and agentic tasks. Yet constructing generalizable synthetic verifiable data remains difficult due to hallucination-prone generation, and weak or trivial verification artifacts that fail to separate strong from weak solutions. Existing approaches often rely on task-specific heuristics or post-hoc filters that do not transfer across domains and lack a principled, universal evaluator of verifiability. In this work, we introduce an evolutionary, task-agnostic, strategy-guided, executably-checkable data synthesis framework that, from minimal seed supervision, jointly synthesizes problems, diverse candidate solutions, and verification artifacts, and iteratively discovers strategies via a consistency-based evaluator that enforces agreement between human-annotated and strategy-induced checks. This pipeline upgrades filtering into principled synthesis: it reliably assembles coherent, verifiable training instances and generalizes without domain-specific rules. Our experiments demonstrate the effectiveness of the proposed approach under both RLVR and model distillation training paradigms. The results show that training with our synthesized data yields significant improvements on both the LiveCodeBench and AgentBench-OS tasks, highlighting the robust generalization of our framework.

Figures (15)

• Figure 1: Overview of EvoSyn, a task-agnostic pipeline for synthesizing verifiable data. From a small human-verified seed data, an evolutionary process discovers a data-filtering strategy via a consistency-based evaluator; this strategy then guides synthesis by generating candidate solutions and tests for new problems, cross-executing them to rank and retain reliable instances while discarding trivial or inconsistent ones. The resulting verifiable dataset (problems, tests, and strong solutions) supports training in diverse tasks.
• Figure 2: Given an initial strategy, the evolutionary algorithm iteratively optimizes it across multiple iterations. Each newly generated strategy is evaluated against our two criteria to determine its effectiveness. The model autonomously explores diverse optimization approaches, ensuring a balance between exploration and exploitation throughout the process.
• Figure 3: The best strategy explored by model on LiveCodeBench.
• Figure 4: Evolutionary process and data-retention trade-off. (a) The evolutionary process consistently discovers stronger strategies, with the best strategy surpassing the initialization by over 10 percentage points within 20 iterations. Score denotes the ratio of seed data instances for which consistency verification is satisfied. (b) Increasing the number of $M$ and $N$ yields more usable, verifiable instances but incurs $O(MN)$ testing execution cost.
• Figure 5: RLVR reward curves comparison across models. EvoSyn-filtered data ($D^{\mathrm{EvoSyn}}$) yields faster, steadier reward growth than random selection ($D^{\mathrm{random}}$).